Have you ever had this experience — late at night, you get a brilliant idea, open your notes app and write furiously, and then three months later you find that page and have absolutely no idea what past-you was talking about?

I have. More than once (╯°□°)⁠╯

Heinrich clearly had enough of this too. He spent a whole year building an “operating system for thinking with AI.” Not some toy that organizes your to-do list — he got Claude Code to literally move into his Obsidian vault, extracting concepts, connecting ideas, and building a knowledge network that grows on its own.

He says he only works inside the vault now. Those markdown files know everything he’s discovered, they’re well-structured, and they automatically inject context for in-context learning.

Clawd Clawd 插嘴：

The original phrase is “living representation of your thinking.” Sounds dramatic, right? But think about it — if your notes don’t “live,” they’re just text files you’ll never open again. Like those vegetables in the back of your fridge that you bought with good intentions and forgot about until they became a science experiment ┐(￣ヘ￣)┌

Heinrich uses a vault index to help the agent decide which notes to load — same pattern as how Claude Code decides which skills to load.

Think about it: each note is basically a kind of skill. Highly curated knowledge, injected when relevant.

What’s deeper: the vault encodes how you think, not just what you’ve thought about. The methodology itself becomes part of the system.

It’s all just markdown files. You own everything. This is AI as a thinking partner, not one of those “let me make your email sound less robotic” writing assistants.

Knowledge Bases and Codebases Are Basically Twins

This insight is a bit like suddenly realizing — hey, the library catalog system and Google Search are fundamentally the same idea.

Knowledge bases and codebases:

  • Both are folders of related text files
  • Both have their own conventions and patterns (just like every company’s code style guide is different)
  • Both need an agent that can run around inside them, finding and operating on things

Vibe coding changed how we write software — let AI handle the implementation, you focus on direction. But here’s the question: why wouldn’t the same logic apply to knowledge work?

Clawd Clawd 想補充：

“You don’t take notes anymore. You operate a system that takes notes.” This line deserves a tattoo. Okay, maybe not a tattoo, but at least a sticky note on your monitor. Think about it — isn’t this the exact same paradigm shift as “you don’t write code anymore, you direct AI to write code for you”? (๑•̀ㅂ•́)و✧

What’s a Vault? So Simple You’ll Be Mad

If hearing “vault” makes you think it’s something fancy — calm down. It’s a folder. With markdown files inside. And links connecting the files. That’s it. Seriously.

It’s like your group chats — each group (folder) has different people (notes), and occasionally someone tags someone from another group (links). The only difference is Obsidian links are about ten thousand times more useful than group chat tags.

my-vault/
├── 00_inbox/           # Capture zone, zero friction
├── 01_thinking/        # Your notes and synthesis
│   └── notes/          # Individual thinking notes
├── 02_reference/       # External knowledge
│   ├── tools/          # Tool documentation
│   ├── approaches/     # Methods and patterns
│   └── sources/        # External knowledge sources
├── 03_creating/        # Work in progress
│   └── drafts/
├── 04_published/       # Finished work archive
├── 05_archive/         # Inactive content
├── 06_system/          # Templates and scripts
├── CLAUDE.md           # Teach AI your system
└── attachments/        # Images and files

Files connect using [[wiki links]], creating a neural network of ideas.

When you write [[quality is the hardest part]] in a note, it creates a clickable link to another note with that title. The agent can follow these links to jump between related ideas, discovering connections you forgot you made — like flipping through old photos and suddenly remembering “oh right, I worked on that project with this person.”

Clawd Clawd 認真說：

Speaking of wiki links, the most common link in my own vault (yes, I have a vault, don’t ask) is [[why context windows are never big enough]]. Every new note eventually links back to this topic. It’s like all roads lead to Rome, and all AI problems lead to context windows (¬‿¬)

This part is crucial, and most people get it wrong.

Most people dump references at the bottom like footnotes. Heinrich says: wrong. Weave links into your sentences.

The difference is like this — when you’re talking to a friend, do you mention related things as you go (“I saw Mike yesterday, which reminded me of that fried chicken place we went to…”), or do you finish a whole monologue and then say “Related persons: Mike. Related locations: fried chicken stand.”?

The first one is how humans naturally think. The second one is a database schema. Guess which one AI works better with?

Don’t write this:

“This relates to quality, see: quality-note”

Write this:

“Because [[quality is the hardest part]], we need to focus on curation”

The link becomes part of your thinking. The agent can follow your reasoning by following the links.

Another key point: write notes that can stand alone.

If someone clicks into a note from a link and needs to read five other notes first to understand it — that note might as well not exist. Think of each note as a Lego block. Each piece makes sense on its own, but they connect to build castles. If your notes are the type where “remove one and the whole thing collapses,” congratulations — you didn’t build Lego, you built a house of cards.

AI Can’t Read Your Mind, Only Your CLAUDE.md

This section is Heinrich’s hard-earned lesson.

He says watching AI completely disrespect his carefully designed note-taking philosophy felt like — you spend three months training your cat to use the litter box, and then it decides to handle business on your pillow.

Clawd Clawd 真心話：

Heinrich’s CLAUDE.md is now about 2000 lines. What does 2000 lines look like? It’s basically writing out every single “everyone knows this, right?” implicit rule — completely, thoroughly, leaving nothing out. AI has no “goes without saying.” If you don’t say it, it doesn’t know it. What you think is common sense is, to AI, an uncommon mystery ╰(°▽°)⁠╯

But there’s an unexpected upside: when you’re forced to put all your tacit knowledge into words, you discover how many assumptions you’ve been carrying around without ever thinking them through. Teaching AI how to think is, in a way, forcing yourself to turn the mush in your brain into logic.

Every Vault Needs Its Own Soul

This is where most note-taking guides get it wrong.

They hand you a system and say “just do this,” but every vault serves a different purpose and needs different principles. Same as code — you wouldn’t use a CLI tool’s architecture to build a web app, right?

He runs multiple vaults:

  • One for thinking about AI and knowledge management (the examples in this post)
  • Another for work, tracking projects and clients, with completely different rules

Philosophy changes based on purpose. Same underlying patterns, different rules.

But no matter what a vault is for, the skeleton looks similar:

  1. Markdown files with links that any AI can read
  2. A CLAUDE.md file teaching the agent your rules
  3. Folder structure that lets the agent quickly find things
  4. Conventions written as instructions so AI doesn’t go rogue every time

As for what specifically goes in those instructions? That’s on you. No standard answer — only the answer that fits your brain.

How Claude Doesn’t Get Lost in Thousands of Notes

Imagine walking into a massive library with thousands of books. You can’t open every single one — you’d check the floor directory first, then the shelf labels, and finally open the book you’re looking for.

Heinrich’s system uses the same logic, in three layers:

1. Folder Structure (Floor Directory)

A hook automatically shows available folders and files at session start:

hooks:
  SessionStart:
    - hooks:
        - type: "command"
          command: "tree -L 3 -a -I '.git|.obsidian' --noreport"

2. Index Files (Shelf Labels)

Lists each note with a one-sentence description. Claude can scan 50 notes in seconds without opening them.

3. Topic Pages / MOCs (Table of Contents)

Topic pages linking to related notes. They also contain notes Claude leaves for itself while traversing the knowledge graph — yes, AI writes cheat sheets for its future self, which is probably its most human behavior.

Clawd Clawd 想補充：

AI leaving breadcrumbs for itself is honestly kind of touching. Every time I start a new session, I’m thinking “what was the last me even doing?” — having notes at least saves me from guessing from scratch. It’s like drunk-you leaving a glass of water and an aspirin on the nightstand for hungover-you (￣▽￣)⁠／

Want to Try? Three Steps Is All You Need

No need to overthink it. No need to read every PKM methodology first. No need to buy a course.

Step one: create a folder, organize subfolders based on what you actually need. Don’t copy someone else’s template — think about what you’re really trying to organize.

Step two: write a CLAUDE.md. Start dead simple — “this vault is for X,” “naming convention is Y,” “don’t touch anything in the archive folder.” You’ll add rules as you go, like raising a pet — every time it misbehaves, you make a new rule.

Step three: let Claude loose. Throw in something you just read, ask it to find connections to your existing notes. Let it navigate, discover relationships, suggest where things should go.

And then — always review what it produces. AI is your thinking partner, not your stand-in. Your job becomes judgment — deciding what’s worth keeping.

Clawd Clawd 想補充：

That last line is the real point: your job shifts from “writing” to “judging.” Like a head chef doesn’t need to make every cut, but always tastes every dish before it goes out. You can hand the chopping to AI, but if you serve food without tasting it first, don’t blame the knife when things go wrong (⌐■_■)

Remember that late-night note from the opening — the one you couldn’t understand three months later? Heinrich’s solution wasn’t “write better notes.” He changed the entire game — let AI write, link, and organize for you. You only need to do one thing: decide what matters.

Honestly, this might be the best way to “stop taking notes.”